-
- GC43F-0864: Interpretable Latent Structure of an AI Emulator of Stochastic Sudden Stratospheric Warmings
-
Board 0864‚ Hall EFG (Poster Hall)NOLA CC
Author(s):Generic 'disconnected' Message
Daniel Boscu, University of Chicago (First Author, Presenting Author)
Daniel Hernandez, University of Chicago
Fabio Ventura, University of Chicago
Dorian Abbot, University of Chicago
Ashesh Chattopadhyay, University of California Santa Cruz
Justin Finkel, The University of Chicago
Pedram Hassanzadeh, Rice University
Sudden Stratospheric Warmings (SSWs) are rare events that disrupt wind speeds in the North Pole’s Stratosphere, or Polar Vortex. Slowing wind speeds in the Polar Vortex are caused by large-scale atmospheric winds. Looking to emulate an idealized polar vortex simulation, we trained a generative Artificial Intelligence model. Our model uses statistical patterns to emulate climate conditions. Conversely, traditional climate models use physical realities to make predictions, resulting in expensive computational costs. After emulating the Polar Vortex, we analyzed the latent space. This architecture component allows the model to generate new data for emulations by sampling from a probability distribution. Plotting the latent space mean in 3D, we discovered that there were well-separated clusters: standard atmosphere and warm atmosphere, or stable and about to transition. This is rarely seen in generative models, which are widely known to be black boxes. The paper introduces an interpretable emulator architecture for climate dynamics with external perturbations, and introduces a latent space probing system for diagnosing how generative AI models internalize rare events. Our findings suggest that effectively built AI emulators can accelerate emulations and uncover physically meaningful patterns in their latent space. Future work includes computing probabilities for SSW and enhancing latent space analysis.
Scientific DisciplineSuggested ItinerariesNeighborhoodType
Enter Note
Go to previous page in this tab
Session
